AI exam 2

machine learning workflow steps

1. data preprocessing

2. model training

3. model evaluation

4. model selection

5. model testing

data sets used for training AI

1. training dataset

2. validation dataset

3. testing dataset

data preprocessing

preparing raw data for model training and analysis including

• data cleaning

• data transformation

• feature engineering

• encoding categorical data

• feature selection

• data splitting

data cleaning

addressing missing values through imputation or removal, correcting errors, and addressing outliers

data transformation

standardizing numerical data to ensure consistency in scale and distribution

feature engineering

creating new, informative features, or modifying existing ones, based on domain knowledge

encoding categorical data

converting non-numerical categories into a numerical format

feature selection

identifying and retaining the most relevant features to reduce data dimensionality

data splitting

dividing the preprocessed data into training, validation, and test sets

model training

focuses on the feeding the training data into machine learning algorithms to train models by tuning parameters like weights to minimize loss

model training process

• feeding a training dataset

• comprising features (input variables)

• labels (output variables)

model training objective

adjusting the parameters of the model so it can accurately map the input to the output, typically involves minimizing a loss function

loss function

quantifies the difference between the predicted outputs of the model and the actual outputs in the training dataho

how to minimize the loss function

optimization algorithms, they iteratively adjust the models parameters like weights in a neural network to minimize the difference between predictions and actual results

model evaluation

assesses the training models on the validation set using relevant metrics for the problem

how is model performance measured

accuracy, precision, recall, F1 score for classification tasks, and MSE and MAE for regression tasks

hyperparameter tuning

a part of the model evaluation process, the configuration settings used to structure the machine learning model to enhance its performance or generalization

examples of hyperparameters

• learning rate

• number of layers in the neural network

• number of trees in a random forest

goal of hyperparameter tuning

find the “sweet spot” where the model is complex enough to capture the underlying patterns in the data, but not so complex that it overfits to the training dataset

parameter vs hyperparameter

a parameter is a variable that is learned from the data during the training process, a hyperparameter is a variable that is set before the training period begins to control the learning process

overfitting

when the model learns the training data too well leading to poor generalization on new data

underfitting

when the model is too simple to capture the underlying patterns in the data, resulting in poor performance on both training sets and test sets

model selection

picking the best validation set performer, if multiple models are similar the simplest model is preferred

goal of model selection

to pick the optimal model for deployment

factors that are used to find the optimal model for deployment

• model simplicity

• training time

• resource requirements

• ease of interpretation

model testing

evaluating the chosen model on an unseen test dataset, which is a representative subset of the data that is expected to show up in the actual environment

dataset splitting goal

developing models that not only perform well on known data but also generalize efficiently to new, unseen data

training dataset

largest portion of the data and is used to train the model, it allows the model to learn patterns and relationships within the data by adjusting its parameters

validation dataset

used to tune hyperparameters and evaluate the model’s performance during training time, helps prevent overfitting and allows for model selection and optimization

test dataset

kept completely separate and is only used to assess the final model’s performance, it helps detect any overfitting and ensures the model will work well when deployed in real-life scenarios

the importance of splitting

1. preventing overfitting

2. model selection, choosing between different models and hyperparameters unbiased

3. unbiased performance estimation

4. iterative improvement, refine the model without contaminating the test dataset

the challenge of overfitting

adjusting the parameters solely based on the training dataset can lead to the model learning not only the general patterns, but also the peculiarities specific to the dataset

how to address overfitting

we continue training the model while monitoring its performance on both the training and validation datasets

what indicates overfitting

errors on the validation dataset may start to increase while errors on the training dataset continue to drop

why do we need a separate test set

1. detects indirect optimization, a subtle form of overfitting

2. information leakage

3. unbiased final evaluation

4. detecting validation set overfitting

5. reliable performance reporting

6. enhancing credibility

neural network

designed to simulate the way a human brain analyzes and processes information

neural network function

consist of layers of interconnected nodes, each of which perform a simple computation; output of these is passed through an activation function which helps standardize the output for the next layer

layers of a neural network

• input layer

• hidden layers

• output layer

input layer

inputs are fed into the network

output layer

`

hidden layer

performs most of the computations, transforming inputs into features the model uses to make predictions; sits between the input and the output, processing data with weights and activation functions to learn patterns

activation functions

determine the output of a node, primarily to introduce non-linearity to the output as most real-world data is non-linear (couldn't handle complex patterns in data)

sigmoid (logistic) function

takes any real-valued number and maps it into a value between 0 and 1, used for model where we need to predict the probability of an output

tanh (hyperbolic tangent) function

maps real-valued numbers to values in the range of -1 to 1, useful when the model needs to predict values that are normalized

ReLU (rectified linear unit) function

allows only positive values to pass through it, and negative values are mapped to zero, allows models to converge faster and reduce the likelihood of vanishing gradient

output layer activation function

chosen according to the specific task, shapes the output into a form that matches the problem statement

regression

type of output layer activation function, no activation function is used or linear if the output is a real-valued prediction

binary classification

type of output layer activation function, the sigmoid function is typically used because it maps predictions to a probability distribution between two classes

multiclass classification

type of output layer activation function, the softmax function is used to produce a probability distribution over multiple classes